Vertical equidistant snakelike route method for monitoring air quality change by using unmanned aerial vehicle

Abstract

The invention provides a vertical equidistant snakelike route method for monitoring air quality change by using an unmanned aerial vehicle, and belongs to the technical field of unmanned aerial vehicles. According to the technical scheme, the multi-unmanned aerial vehicle intelligent cooperative monitoring method for refined air quality management is constructed by fusing advanced technologies such as a knowledge graph, deep reinforcement learning, federal learning, a graph neural network and a generative adversarial network. According to the method, multi-source heterogeneous data are fully utilized, the unmanned aerial vehicle route and the sampling strategy are adaptively optimized, and the monitoring efficiency and precision are remarkably improved. Meanwhile, the generated high-quality air quality distribution diagram visually reflects the regional pollution condition, and visual decision support is provided for precise pollution control. The invention provides a new thought and method for the fields of wisdom environmental protection, pollution traceability and the like, and has a wide application prospect.

Description

Technical Field

[0001] The present invention relates to the technical field of drones, and specifically to a vertical equidistant serpentine route method for monitoring air quality changes using drones. Background Technique

[0002] With the continuous acceleration of the industrialization and urbanization processes, the problem of air pollution has become increasingly severe, posing a serious threat to human health and the ecological environment. To effectively address this challenge, accurately monitoring air quality changes and promptly detecting and warning pollution events have become an urgent need for environmental management departments. Traditional air quality monitoring mainly relies on ground-based fixed stations, which have deficiencies such as sparse distribution points and limited coverage, and it is difficult to meet the requirements of refined management. Therefore, there is an urgent need to develop a new monitoring method that is flexible, efficient, highly accurate, and has a wide coverage.

[0003] Currently, commonly used air quality monitoring technologies mainly include ground monitoring stations, satellite remote sensing, lidar, etc. Ground monitoring stations measure the concentration of pollutants in the air directly by deploying fixed measurement points and using various sensors, and have high accuracy and continuity. However, limited by the construction cost and site selection conditions, the monitoring stations are usually relatively sparse and it is difficult to comprehensively reflect the regional pollution distribution. Satellite remote sensing uses a spectral imager carried by a satellite to indirectly estimate the surface pollution status by retrieving parameters such as aerosol optical thickness in the atmosphere. However, affected by cloudy and rainy weather, the time resolution of satellite remote sensing is low, and it is difficult to obtain pollution information in the near-surface layer. Lidar analyzes the backscattered light signal to retrieve the atmospheric aerosol profile, but the equipment is expensive and not sensitive to near-surface data. In summary, traditional technologies have deficiencies in terms of monitoring efficiency, accuracy, and cost, and there is an urgent need for innovative solutions.

[0004] In recent years, with the rapid development of unmanned aerial vehicle (UAV) technology and the continuous improvement of sensor performance, using UAVs for environmental monitoring has become a research hotspot. On the one hand, light and small UAVs have advantages such as flexibility, mobility, and low cost, enabling rapid coverage and data collection of the target area. On the other hand, the emergence of integrated and miniaturized sensors has made it possible for UAVs to carry multi-parameter monitoring equipment. In addition, in terms of information fusion and knowledge mining, artificial intelligence technologies such as knowledge graphs and deep reinforcement learning have made great progress. Knowledge graphs represent and reason about knowledge in complex scenarios by constructing semantic association networks between concepts; deep reinforcement learning enables autonomous decision-making of agents in dynamic environments through end-to-end policy optimization; federated learning enables collaborative learning among multiple participants while protecting data privacy through encrypted parameter aggregation. These technologies have injected new vitality into the field of environmental monitoring and provided new ideas for solving traditional technical problems. Although UAV technology has made great progress in recent years, there is currently no mature technology that can be used as a reference for achieving automated and even intelligent environmental monitoring using UAVs. In this context, how to design the control method of UAVs around the purpose of environmental monitoring is a technical problem to be solved in this field. Summary of the Invention

[0005] The technical problem to be solved by the present invention is: how to design a UAV control method that can efficiently and accurately monitor air quality changes.

[0006] To achieve the above technical objectives, the present invention adopts the following technical solutions:

[0007] A vertical equidistant serpentine flight path method for using UAVs to monitor air quality changes, comprising:

[0008] 1) Collect regional environmental status information, introduce knowledge graph and geographic information system technologies to construct an environmental status knowledge base, and realize semantic representation and reasoning of regional environmental information as the input for UAV flight path planning;

[0009] 2) Use a hierarchical deep reinforcement learning algorithm to adaptively optimize the flight path of each UAV, and dynamically adjust the monitoring height and distance through hierarchical decision-making of global path planning and local trajectory optimization;

[0010] 3) Construct a decentralized multi-UAV collaborative monitoring framework based on federated learning, and regard each UAV as a node in the graph;

[0011] 4) Use an adaptive graph convolutional network enhanced by a graph attention network to achieve information fusion and collaborative decision-making between UAV nodes, and dynamically adjust the neighborhood information aggregation method while considering node features and topological structures;

[0012] 5) The UAV executes the air quality monitoring task according to the optimized flight path and cooperation strategy, adopts a multi-modal data-driven adaptive sampling strategy, and dynamically adjusts the sampling density and frequency according to the collected data and prior knowledge; draws a serpentine flight path based on the diagonal length of the monitored area, and conducts dot monitoring at the same distance at different heights;

[0013] 6) Use a gated attention convolutional long short-term memory network to fuse the monitoring data, adaptively control the importance of data at different heights and distances through the gated unit, and extract multi-scale spatio-temporal features;

[0014] 7) Introduce a conditional attention generative adversarial network to generate an air quality distribution map based on the fused features.

[0015] Preferably, in step 2), the UAV hovers and monitors at heights that are integer multiples of the unit distance from the ground, stays in the air for the unit time for each increase in the unit distance, and records the different air index values during this time.

[0016] Preferably, the unit distance is 10 meters; the unit time is 1 min.

[0017] Preferably, in step 1), introduce a knowledge graph and geographic information system technology to construct an environmental state knowledge base: through ontological modeling and semantic association of multi-source heterogeneous data including environmental monitoring data, geospatial data, and meteorological data, form a semantic network including concept nodes and relationship edges to realize the semantic representation of regional environmental information; at the same time, use GIS technology to conduct spatial visualization analysis of environmental elements.

[0018] Preferably, in step 1), let the environmental state feature vector be The set of concept nodes in the knowledge base is The set of relationship edges is Then the regional environmental state knowledge base is represented as a weighted directed graph

[0019] Preferably, based on the constructed environmental state knowledge base, adopt a hierarchical deep reinforcement learning algorithm to adaptively optimize the UAV flight path; the global path planning layer takes the regional environmental characteristics as the observation state, uses data including coverage rate and energy consumption as the reward function, and solves the optimal cruise path through the deep Q network; the local trajectory optimization layer, under the global path constraint, takes the UAV's own state as the observation value, uses data including flight path smoothness and sampling accuracy as the reward function, and solves the optimal trajectory parameters through the proximal policy optimization algorithm.

[0020] Preferably, in the process of the adaptive optimization, let the state of the kth UAV be where is the position, is the speed, are the Euler angles; the global path planning strategy is the local trajectory optimization strategy is then the adaptive optimization model of the UAV flight path is expressed as:

[0021]

[0022] where, τ g and τ l respectively represent the global path and the local trajectory, γ ∈ [0, 1] is the discount factor, and are the global reward and the local reward respectively, Δt is the discrete time step, is the acceleration, P is the feasible flight area, v max and a max are the maximum speed and the maximum acceleration respectively.

[0023] Preferably, in step 3), a decentralized multi-UAV collaborative monitoring framework based on federated learning is constructed: each UAV realizes information fusion and policy iteration among nodes through an adaptive graph convolutional neural network enhanced by a graph attention network based on local observations and decision-making experiences.

[0024] Preferably, let the parameters of the i-th UAV be Wi, its set of neighbor nodes be N(i), and the attention weight be aij, then the local policy gradient is expressed as:

[0025]

[0026] where, hi is the feature representation of the i-th UAV, W and a are the attention parameters; under the federated learning framework, each UAV realizes the global optimization of the policy through encrypted gradient aggregation:

[0027]

[0028] where, M is the number of UAVs, η is the learning rate, ni is the number of samples of the i-th UAV, and n is the total number of samples;

[0029] When performing the monitoring task, each UAV plans the sampling flight path in a diagonal snake-like manner, and adaptively adjusts the sampling density and frequency according to the multi-scale spatio-temporal features extracted by the gated attention convolutional long short-term memory network; let be the monitoring data collected by the i-th UAV at time t, H and W are the height and width of the data matrix respectively, then the forward propagation process of the gated attention convolutional long short-term memory network is:

[0030]

[0031] Among them, are the forget gate, input gate, and output gate respectively, is the cell state, is the hidden state, W and b are learnable parameters, σ is the sigmoid activation function, and ⊙ is the Hadamard product; the gated unit realizes the adaptive optimization of the sampling density and frequency by dynamically adjusting the weights of different-scale features.

[0032] Preferably, in step 7), a conditional attention generative adversarial network is introduced: the generator G takes the fused multi-scale spatio-temporal feature Z and random noise ∈ as inputs and generates a realistic distribution map I g = G(z, ∈); the discriminator D takes the real distribution map Ir or the generated distribution map Ig as inputs and outputs the real probability D(I); the generator and the discriminator are continuously optimized through the minimax game and finally reach the Nash equilibrium; the objective function is expressed as:

[0033]

[0034] where p data (I r ) is the distribution function of the real distribution map, and p ∈ (∈) is the distribution function of the random noise.

[0035] The present invention has the following two key innovation points:

[0036] First, an adaptive optimization of the UAV flight path is achieved by using a hierarchical deep reinforcement learning algorithm. Different from the traditional fixed flight path planning method, this algorithm dynamically adjusts the flight attitude and sampling strategy of the UAV through hierarchical decision-making of global path planning and local trajectory optimization, combines the regional environmental characteristics and the UAV's own state, minimizes the energy consumption while ensuring the coverage rate, and significantly improves the monitoring efficiency.

[0037] Second, the idea of federated learning is introduced to construct a decentralized multi-UAV collaborative learning framework. The traditional centralized collaborative control method relies on a central node and is prone to single-point failures and communication bottlenecks. However, the present invention realizes distributed information fusion and policy optimization through a graph neural network, protects data privacy, reduces the communication load, enhances the robustness and scalability of the system. At the same time, the proposed gated attention convolutional long short-term memory network can adaptively extract multi-scale spatio-temporal features, further guide the dynamic optimization of the sampling density and frequency, and improve the monitoring accuracy.

[0038] The present invention provides a vertical equidistant serpentine route method for monitoring air quality changes using unmanned aerial vehicles (UAVs). Through the integration of cutting-edge technologies such as knowledge graphs, deep reinforcement learning, federated learning, graph neural networks, and generative adversarial networks, this technical solution constructs a multi-UAV intelligent collaborative monitoring method for refined air quality management. This method makes full use of multi-source heterogeneous data, adaptively optimizes UAV routes and sampling strategies, and significantly improves monitoring efficiency and accuracy. At the same time, the generated high-quality air quality distribution map intuitively reflects the regional pollution status, providing visual decision support for precise pollution control. The present invention provides new ideas and methods for the fields of intelligent environmental protection, pollution source tracing, etc., and has broad application prospects.

[0039] The present invention uses UAVs equipped with special sensors to conduct three-dimensional dynamic scans of the target area along a preset route, obtaining air quality data at different heights and positions, and comprehensively reflecting the regional pollution status. This innovative method breaks through the limitations of traditional ground monitoring, greatly improves monitoring efficiency and accuracy, provides key technical support for precise pollution control, environmental emergencies, etc., and has important scientific significance and application value.

[0040] Based on the latest progress in UAV monitoring and artificial intelligence technologies, the present invention proposes an innovative vertical equidistant serpentine route method. By carefully designing flight routes, optimizing the combination of on-board sensors, and introducing cutting-edge technologies such as knowledge graphs, deep reinforcement learning, and federated learning, the flexibility, accuracy, and intelligence level of UAV monitoring are significantly improved. At the same time, the constructed refined air quality management method provides a new technical means for scientific decision-making by environmental protection departments, and is of great significance for promoting air pollution prevention and control and improving regional environmental quality. The research results of the present invention are expected to be widely applied and promoted in the fields of intelligent environmental protection, UAV applications, etc.

[0041] Compared with the prior art, the present invention has the following beneficial effects:

[0042] 1. Significantly improve monitoring efficiency and accuracy. By integrating cutting-edge technologies such as knowledge graphs, deep reinforcement learning, and federated learning, the present invention makes full use of multi-source heterogeneous data to optimize UAV routes and sampling strategies. On the one hand, the hierarchical reinforcement learning algorithm based on the regional environmental state knowledge base can adaptively plan global paths and local trajectories, reducing manual operations and improving monitoring coverage and efficiency. On the other hand, the multi-UAV collaborative learning mechanism and gated attention network can dynamically adjust the sampling density and frequency, and combined with the serpentine diagonal track, achieve high-resolution, all-round three-dimensional monitoring, greatly improving data accuracy and reliability.

[0043] 2. Provide visual decision-making support for precise pollution control. The present invention uses a conditional generative adversarial network to fuse multi-scale spatio-temporal features to generate high-quality air quality distribution maps. Compared with traditional scatter interpolation methods, the generated distribution maps have higher resolution, stronger continuity, and finer texture features, and can intuitively reflect the dynamic changes in regional pollution conditions. This provides intuitive and reliable data support and visualization tools for air pollution source tracing, key area identification, pollution control decision-making, etc., and helps with precise pollution control. The environmental protection department can adjust the supervision strategy in a timely manner according to the air quality distribution map, rationally allocate pollution control resources, and minimize the pollution hazard to the greatest extent.

[0044] In summary, the present invention provides an intelligent and efficient technical solution for the refined monitoring and management of the atmospheric environment, meets the development needs of current intelligent environmental protection and precise pollution control, and has broad application prospects and promotion value. Detailed implementation manners

[0045] The following will describe the detailed implementation manners of the present invention in detail. To avoid excessive unnecessary details, the well-known structures or functions will not be described in detail in the following embodiments. The approximate language used in the following embodiments can be used for quantitative expressions, indicating that certain changes in quantity are allowed without changing the basic function. Unless otherwise defined, the technical and scientific terms used in the following embodiments have the same meaning as commonly understood by those skilled in the art to which the present invention belongs.

[0046] A vertical equidistant serpentine route method for monitoring air quality changes using an unmanned aerial vehicle, comprising the following steps:

[0047] (1) Collect the environmental status information of the collection area, introduce knowledge graph and geographic information system technologies to build an environmental status knowledge base, realize the semantic representation and reasoning of regional environmental information, and use it as the input for UAV route planning; (2) Use the hierarchical deep reinforcement learning algorithm to adaptively optimize the route of each UAV, and through the hierarchical decision-making of global path planning and local trajectory optimization, dynamically adjust the monitoring height and distance. The UAVs hover and monitor at heights of 10 meters, 20 meters, 30 meters, etc. above the ground respectively. For every 10-meter rise, they stay in the air for 1 minute, and record the different air index values during this time; (3) Build a decentralized multi-UAV collaborative monitoring framework based on federated learning, and regard each UAV as a node in the graph; (4) Use the adaptive graph convolutional network enhanced by the graph attention network to realize information fusion and collaborative decision-making among UAV nodes, and at the same time consider node features and topological structures, and dynamically adjust the way of aggregating neighborhood information; (5) The UAVs perform air quality monitoring tasks according to the optimized routes and collaborative strategies, adopt a multi-modal data-driven adaptive sampling strategy, and dynamically adjust the sampling density and frequency according to the collected data and prior knowledge. Draw a serpentine route based on the diagonal length of the monitored area, and perform dot monitoring at the same distance at different heights to improve accuracy; (6) Use a gated attention convolutional long short-term memory network to fuse the monitoring data, and adaptively control the importance of data at different heights and distances through the gated unit to extract multi-scale spatio-temporal features; (7) Introduce a conditional attention generative adversarial network to generate a high-resolution, continuous and smooth air quality distribution map based on the fused features, providing decision support for refined air quality management.

[0048] The present invention proposes a multi-UAV intelligent collaborative monitoring method for refined air quality management. One of the key steps is to use multi-source heterogeneous data fusion and adaptive sampling strategies to achieve efficient and accurate air quality monitoring. Specifically, this step can be refined into the following aspects:

[0049] First of all, the present invention introduces knowledge graph and geographic information system (GIS) technologies to build a regional environmental status knowledge base. Through the ontological modeling and semantic association of multi-source heterogeneous data such as environmental monitoring data, geospatial data, and meteorological data, a semantic network with concept nodes as the core and relationship edges as the link is formed to realize the semantic representation of regional environmental information. At the same time, use GIS technology to perform spatial visualization analysis of environmental elements to reveal the law of pollutant diffusion. Let the environmental status feature vector be The set of concept nodes in the knowledge base is The set of relationship edges is Then the regional environmental status knowledge base can be represented as a weighted directed graph Secondly, based on the constructed environmental state knowledge base, a hierarchical deep reinforcement learning algorithm is used to adaptively optimize the UAV flight path. The global path planning layer takes the regional environmental features as the observation state, and uses the coverage rate, energy consumption, etc. as the reward function to solve the optimal cruise path through the Deep Q-Network (DQN). Under the global path constraint, the local trajectory optimization layer takes the UAV's own state (position, speed, power, etc.) as the observation value, and uses the track smoothness, sampling accuracy, etc. as the reward function to solve the optimal track parameters (speed, acceleration, heading, etc.) through the Proximal Policy Optimization (PPO) algorithm. Let the state of the k-th UAV be where is the position,[[]] is the speed,[[]] is the Euler angle. The global path planning strategy is The local trajectory optimization strategy is Then the UAV flight path adaptive optimization model can be expressed as:

[0050]

[0051] where, τ g and τ l represent the global path and the local trajectory respectively, γ ∈ [0, 1] is the discount factor, and are the global reward and the local reward respectively, Δt is the discrete time step, is the acceleration, P is the feasible flight area, v max and a max are the maximum speed and the maximum acceleration respectively.

[0052] In the multi-UAV cooperative monitoring framework, the present invention designs a decentralized multi-agent cooperative learning mechanism based on the idea of federated learning. Each UAV realizes information fusion and policy iteration among nodes through an Adaptive Graph Convolutional Neural Network Enhanced by Graph Attention Network (GAT-AdaGCN) based on local observations and decision-making experience. Let the parameters of the i-th UAV be Wi, its set of neighbor nodes be N(i), and the attention weight be aij, then the local policy gradient can be expressed as:

[0053]

[0054] where, hi is the feature representation of the i-th UAV, and W and a are the attention parameters. Under the federated learning framework, each UAV realizes the global optimization of the policy through encrypted gradient aggregation:

[0055]

[0056] where, M is the number of UAVs, η is the learning rate, ni is the number of samples of the i-th UAV, and n is the total number of samples.

[0057] When performing the monitoring task, each UAV plans the sampling route in a diagonal serpentine manner, and adaptively adjusts the sampling density and frequency according to the multi-scale spatio-temporal features extracted by the gated attention convolutional long short-term memory network (GA-ConvLSTM). Let be the monitoring data collected by the i-th UAV at time t, and H and W be the height and width of the data matrix respectively. Then the forward propagation process of GA-ConvLSTM is as follows:

[0058]

[0059] Among them, are the forget gate, input gate and output gate respectively, is the cell state, is the hidden state, W and b are learnable parameters, σ is the sigmoid activation function, and ⊙ is the Hadamard product. The gated unit realizes the adaptive optimization of the sampling density and frequency by dynamically adjusting the weights of different scale features.

[0060] Finally, the present invention uses a conditional generative adversarial network (CGAN) to generate a high-quality air quality distribution map. The generator G takes the fused multi-scale spatio-temporal feature Z and random noise ∈ as inputs and generates a realistic distribution map I g = G(z, ∈). The discriminator D takes the real distribution map Ir or the generated distribution map Ig as inputs and outputs the real probability D(I). The generator and the discriminator are continuously optimized through the minimax game and finally reach the Nash equilibrium. The objective function can be expressed as:

[0061]

[0062] Among them, p data (I r ) is the distribution function of the real distribution map, and p ∈ (∈) is the distribution function of the random noise.

[0063] The above has described the embodiments of the present invention in detail, but the content described is only the preferred embodiments of the present invention and is not used to limit the present invention. Any modifications, equivalent replacements and improvements made within the scope of the application of the present invention shall be included within the protection scope of the present invention.

Claims

1. A vertical equidistant serpentine flight method for monitoring air quality changes using unmanned aerial vehicles, characterized in that: include: 1) Collect regional environmental status information, introduce knowledge graph and geographic information system technology to build an environmental status knowledge base, realize the semantic representation and reasoning of regional environmental information, and use it as the input for UAV route planning; 2) Adaptively optimize the route of each drone using a hierarchical deep reinforcement learning algorithm, dynamically adjust the monitoring height and distance through hierarchical decision-making of global path planning and local trajectory optimization; 3) Construct a decentralized multi-UAV collaborative monitoring framework based on federated learning, treating each UAV as a node in the graph; 4) Using the adaptive graph convolutional network enhanced by the graph attention network to realize information fusion and collaborative decision-making among drone nodes, while considering the node characteristics and topological structure, and dynamically adjusting the neighborhood information aggregation method; 5) The drone performs air quality monitoring tasks according to the optimized route and collaborative strategy, adopts a multimodal data-driven adaptive sampling strategy, and dynamically adjusts the sampling density and frequency according to the collected data and prior knowledge; a serpentine route is drawn based on the diagonal length of the monitored area, and point monitoring is performed at different heights and the same distance; 6) A gated attention convolutional long short-term memory network is used to fuse the monitoring data, and the importance of data at different heights and distances is adaptively controlled through the gated unit to extract multi-scale spatial-temporal features; 7) Introduce the conditional attention generative adversarial network to generate an air quality distribution map based on the fused features.

2. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 1, characterized in that: In step 2), the UAV performs hovering monitoring at an integer unit distance from the ground, stays in the air for a unit time every time it rises a unit distance, and records different air index values ​​during this time.

3. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 2, characterized in that: The unit distance is 10 meters; the unit time is 1 minute.

4. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 1, characterized in that: In step 1), knowledge graph and geographic information system technology are introduced to construct an environmental status knowledge base: by ontological modeling and semantic association of multi-source heterogeneous data including environmental monitoring data, geospatial data, and meteorological data, a semantic network including concept nodes and relationship edges is formed to achieve semantic representation of regional environmental information; at the same time, GIS technology is used to perform spatial visualization analysis of environmental elements.

5. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 4, characterized in that: In step 1), let the environmental state feature vector be The concept node set in the knowledge base is The relationship edge set is The regional environmental state knowledge base is represented as a weighted directed graph 6. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 5, characterized in that: Based on the constructed environmental state knowledge base, a hierarchical deep reinforcement learning algorithm is used to adaptively optimize the UAV route; the global path planning layer takes the regional environmental characteristics as the observation state, and data including coverage and energy consumption as the reward function, and solves the optimal cruise path through a deep Q network; under the global path constraint, the local trajectory optimization layer takes the UAV's own state as the observation value, and data including track smoothness and sampling accuracy as the reward function, and solves the optimal track parameters through a proximal strategy optimization algorithm.

7. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 6, characterized in that: In the process of the adaptive optimization, the state of the kth UAV is assumed to be in For location, For speed, is the Euler angle; the global path planning strategy is The local trajectory optimization strategy is The UAV route adaptive optimization model is expressed as: Among them, τ g and τ l denote the global path and local trajectory respectively, γ∈[0,1] is the discount factor, and are global reward and local reward respectively, Δt is the discrete time step, is the acceleration, P is the feasible flight area, v max and a max are the maximum speed and maximum acceleration respectively.

8. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 1, characterized in that: In step 3), a decentralized multi-UAV collaborative monitoring framework based on federated learning is constructed: each UAV realizes information fusion and strategy iteration between nodes through an adaptive graph convolutional neural network enhanced by a graph attention network based on local observation and decision-making experience.

9. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 8, characterized in that: Assume that the parameters of the i-th drone are Wi, its neighbor node set is N(i), and the attention weight is aij, then the local policy gradient is expressed as: Where hi is the feature representation of the i-th drone, W and a are attention parameters; under the federated learning framework, each drone achieves global optimization of the strategy through encrypted gradient aggregation: Among them, M is the number of drones, η is the learning rate, ni is the number of samples of the i-th drone, and n is the total number of samples; When performing monitoring tasks, each drone plans the sampling route in a diagonal serpentine manner, and adaptively adjusts the sampling density and frequency according to the multi-scale spatiotemporal features extracted by the gated attention convolutional long short-term memory network; is the monitoring data collected by the i-th drone at time t, H and W are the height and width of the data matrix respectively, then the forward propagation process of the gated attention convolutional long short-term memory network is: in, They are forget gate, input gate and output gate respectively. is the cell state, is the hidden state, W and b are learnable parameters, σ is the sigmoid activation function, and ⊙ is the Hadamard product. The gated unit achieves adaptive optimization of sampling density and frequency by dynamically adjusting the weights of features of different scales.

10. The vertical equidistant serpentine flight method for monitoring air quality changes using an unmanned aerial vehicle according to claim 1, characterized in that: In step 7), the conditional attention generative adversarial network is introduced: the generator G takes the fused multi-scale spatiotemporal features Z and random noise ∈ as input to generate a realistic distribution map I g =G(z,∈); the discriminator D takes the real distribution map Ir or the generated distribution map Ig as input and outputs the real probability D(I); the generator and the discriminator are continuously optimized through the minimax game and finally reach the Nash equilibrium; the objective function is expressed as: Among them, p data (I r ) is the distribution function of the true distribution graph, p ∈ (∈) is the distribution function of random noise.

Images